ML & DL Prediction Suite · 9 Models Live

Zanevo

Run real predictions. Live.

A unified platform for production-trained ML & DL models across healthcare, business, real estate,education, and personality. Enter inputs, get predictions backed by real datasets.

Step 01Select a model

Step 02Enter input data

Step 03Get your prediction

Models deployed

Across healthcare, business,
real estate & education

4 Models · Classification

2 Models
Regression

1 Model
Clustering

2 Models
Neural Net

All systems live

All Models

9 live

9 Text Features

Numpy Neural Net · MBTI Kaggle Dataset

MBTI Personality Predictor

Predicts a 4-letter personality type from writing style — pronoun usage, punctuation, sentence length. Four binary classifiers built from scratch, no sklearn or PyTorch.

Take the quiz

75% acc

U-Net (from scratch) · Plant Disease Segmentation

Plant Disease Detection

Segments the exact diseased region on a leaf photo and identifies which of 15 diseases it is, not just a sick/healthy label.

Analyze leaf

Silhouette 0.4129

K-Means · Olist E-Commerce Dataset

Customer Segmentation

Segments 83,000 customers into 3 behavioral groups using RFM analysis. K=3 selected via Elbow and Silhouette methods; Frequency dropped after showing zero variance.

Run segmentation

R² 0.9315

Ridge Regression · Ames Housing

Ames House Price

Predicts residential sale prices from 80+ features. Ridge selected over LightGBM for better out-of-box generalization. MAE of $11,004 on the test set.

Run prediction

F1 0.72

XGBoost · UCI Adult Census

Adult Census Income

Classifies income above or below $50K on an imbalanced dataset (76/24 split). Threshold shifted from 0.5 → 0.4 to improve recall on the minority class.

Run assessment

Threshold 0.30

XGBoost · IBM HR Dataset

Employee Attrition

Predicts employee flight risk from satisfaction, tenure, and compensation. Threshold lowered to 0.30 to prioritize early detection over precision.

Run assessment

Threshold 0.323

Random Forest · Telco Dataset

Customer Churn

Identifies telecom customers at cancellation risk. Threshold tuned below 0.5 to maximize recall, catching more churners at the cost of some precision.

Run assessment

13 Features

Logistic Regression · Cleveland Dataset

Heart Disease Risk

Evaluates cardiovascular risk from ECG results, cholesterol, and chest pain type. StandardScaler pipeline on the UCI Cleveland Heart Disease dataset.

Run assessment

R² ≈ 0.99

Random Forest · Student Performance Prediction

Student Performance

Predicts a student's performance index from study hours, previous scores, and sleep. Previous scores dominate feature importance by a wide margin.

Run prediction

Technical Decisions

Why I built it this way

These aren't just demos. Each model involved a deliberate tradeoff, between precision and recall, between model complexity and generalization, between a good metric and a useful one.

Dice loss fixed what cross-entropy hid

The U-Net's training loss dropped smoothly and looked healthy on cross-entropy alone, but per-class IoU told a different story: several disease classes sat at or near zero while background pixels carried the average. Switching to combined cross-entropy + Dice loss, which scores region overlap directly rather than averaging pixel by pixel, improved every one of the 15 classes. The strongest jump was Coffee Leaf Rust: 0.010 → 0.418.

Held-out test accuracy decides trust, not validation alone

Validation scores made all four MBTI dimensions look roughly usable. Re-checking against a completely unseen test set told a different story: T/F held up (0.613 vs a 0.541 baseline), but I/E, N/S, and J/P never cleared their own baseline. Only one of four dimensions earned its confidence score.

Threshold tuning on imbalanced classifiers

Churn (0.323), Attrition (0.30), and Census Income (0.4) all use custom thresholds below 0.5. The default threshold optimizes accuracy, the wrong objective when missing a churner costs more than a false alarm. Lowering the threshold shifts the tradeoff toward catching more positives.

Ridge over LightGBM for Ames Housing

LightGBM showed a 6.4-point train/test gap (0.9826 vs 0.9183) without tuning, a clear sign of overfitting. Ridge with L2 regularization handled the 167-feature post-encoding space better out of the box. A well-regularized linear model beat the tree here.

Dataset size and interpretability drive algorithm choice

Heart Disease uses Logistic Regression because the Cleveland dataset is only 303 rows and clinical interpretability matters more than marginal accuracy gains. Attrition uses XGBoost because HR data has complex feature interactions that linear models structurally miss. Algorithm selection followed the data.

Feature engineering validated by importance scores

Six features were engineered for the Attrition model including IsOverworked, combining overtime status with poor work-life balance as a stress signal. It ranked in the top 5 most important features out of 35, confirming the engineering decision rather than just adding noise.

F1 over accuracy on imbalanced datasets

The UCI Adult dataset is 76/24 class-imbalanced. Predicting everyone as ≤$50K hits 76% accuracy while being completely useless. The IBM HR dataset is 84/16. In both cases F1 and recall are the meaningful metrics, accuracy is a number that hides a broken model.

Dropping zero-variance features before clustering

In the Olist segmentation project, Frequency showed that over 90% of customers had placed exactly one order. After log transform and outlier removal it had a standard deviation of 0.0, meaning it could not separate any customers from each other. A feature that adds no signal only distorts centroid positions. It was dropped.